Kernel: KVM uaccess hardening, deadline accounting, and lockless page cache

The day's kernel traffic centers on KVM teardown safety and scheduler bandwidth accounting, with lockless page-cache and BPF verifier work close behind. List hygiene and library API fixes also drew maintainer attention ahead of 7.3.

KVM hardens uaccess against a dying VM

KVM maintainer posted a 23-patch v4 series that fixes user-access corruption of unrelated processes during VM teardown, including nested VMX, and adds hardening against bad uaccess. The bug matters because a dying guest can still touch kernel paths that copy to or from user memory belonging to other tasks. Review is underway on linux-kernel.

BPF verifier scopes callback arguments

A six-patch bpf-next v4 series scopes callback arguments to their frame so ringbuf and map helpers cannot leak kernel stack contents. The change is a verifier fix rather than a new helper surface. It closes a class of leaks that unprivileged or semi-trusted BPF programs could otherwise observe.

Deadline bandwidth accounting gets a large fix series

A 29-patch v5 series addresses 25 SCHED_DEADLINE bandwidth accounting bugs, three of which can crash, and adds selftests. Six participants have discussed it across 78 messages, with some overlap among the fixes noted in review. Correct accounting is what keeps deadline tasks from oversubscribing CPU bandwidth.

RFC lets filemap and shmem reads skip folio references

Meta posted RFC v2, originating with Linus, that introduces PG_rcu_free and an RCU grace period so filemap and shmem can be read without taking folio references. The goal is to avoid refcount contention on the page cache under lockless readers. The series is still a proposal, not a merge candidate.

Library fixes for 7.3 reopen kvmalloc GFP rules

A four-patch series of library fixes for 7.3 drew a debate among Linus and the mm maintainers over kvmalloc GFP_ATOMIC and vmalloc support, after an rhashtable deadlock and maple_tree RCU fixes. The question is which allocation modes kvmalloc may safely advertise. That choice affects every caller that falls back from kmalloc to vmalloc under memory pressure.

get_maintainer would stop copying LKML on every patch

A two-patch v3 changes get_maintainer.pl so LKML is only a fallback, not an address added to every submission. The intent is to cut list volume without dropping patches that have no other list. Four participants have commented on the v3.

Perf hashmaps drop aliasing UB and ERR_PTR

A 26-patch series migrates perf to type-safe DECLARE_HASHMAP, removing strict-aliasing undefined behavior and ERR_PTR usage. The work is posted as v1 with a single author so far. It is an internal cleanup of perf's hashmap rather than a user-visible ABI change.

fmod_ret proposed for thousands of error-returning functions

Alexei Starovoitov posted v2 of a series that expands fmod_ret so it can attach to more than 12,000 functions that return errors, using a build-time scan and untrusted pointers. The mechanism lets BPF modify return values on those functions. Review is on the bpf list.